Backfill backend_type for existing agents during startup reconciliation

[geoffjay]

Summary

Existing agents in the database have incorrect backend_type values (hardcoded
"tmux" from before the fix). The startup reconciliation loop should detect and
correct stale backend_type values so that agents created before the fix are
reported accurately.

Context

AgentManager::reconcile() in crates/orchestrator/src/manager.rs (line 393)
already iterates over all running agents at startup and checks their session
health. This is the natural place to also fix stale backend_type values.

After #1236 lands, newly spawned agents will have the correct backend_type.
But agents that were persisted before the fix will still have "tmux" (or
None) in the database. This issue handles the data migration aspect.

Acceptance Criteria

• In reconcile_agent(), check if agent.backend_type differs from
  self.backend.backend_name() and update it if so
• Log a warning when correcting a stale backend_type value (include
  agent ID, old value, new value)
• Persist the corrected value via self.storage.update(&agent)
• Only update agents whose sessions are confirmed alive (don't touch
  stopped/failed agents that might have legitimately run on a different backend)
• cargo clippy and cargo fmt pass

Key Files

• crates/orchestrator/src/manager.rs -- reconcile_agent() method (line 434)

Blocked By

• Set correct backend_type in AgentManager spawn_agent and restart_agent #1236 (the core fix must land first so the reconciliation uses the same
  backend_name() source of truth)

Stack Base

Stack on: feature/autonomous-pipeline
Blocked by: #1236

[geoffjay]

✅ Triage Complete

Field	Value
Type	bug
Complexity	small
Milestone	v0.3.0 — Stabilization & Code Quality
Labels	bug, complexity:small

Notes:

• No duplicates found. The backend_type series (Add backend_name() method to ExecutionBackend trait #1235–Add tests for backend_type propagation across agent lifecycle #1239) are all distinct; this issue owns the startup reconciliation data-correction path only.
• bug retained — existing agents in the DB carry stale "tmux" values that misrepresent their actual backend, which is incorrect data.
• Complexity small: all changes are within reconcile_agent() in a single file (crates/orchestrator/src/manager.rs, line 434). The logic is fully prescribed by the acceptance criteria: compare agent.backend_type against self.backend.backend_name(), log a warning with old/new values, call self.storage.update(&agent), and gate on session-alive status. Estimated ~20–30 lines.
• Milestone v0.3.0: data integrity correction for a known bug — fits the stabilization goal exactly.
• Blocked by Set correct backend_type in AgentManager spawn_agent and restart_agent #1236 — reconciliation should use the same backend_name() source of truth that the fix introduces.
• Test coverage for this reconciliation path is tracked separately in Add tests for backend_type propagation across agent lifecycle #1239 (acceptance criterion 3).
• Acceptance criteria are complete and unambiguous; no clarification needed.